ACAT-V for Rare-Variant Set Testing

Why Combine Variant-Level Evidence?

A rare-variant set may contain many variants, while association evidence is concentrated in only one or a few of them. A burden test can lose power when effects point in different directions, because its aggregate signal may cancel. A variance-component test such as SKAT instead aggregates quadratic score evidence across variants. ACAT-V takes a different route: it combines the variant-level p-values themselves, using a transformation that gives very small p-values substantial influence.

The method is useful when the set-level question is whether there is evidence of association somewhere in a prespecified set, without requiring that every variant contribute similarly. It does not, by itself, determine which variant is causal, estimate a set-wide effect, or protect against poorly calibrated variant-level tests.

Definition: For a set of $m$ variants, let $p_j$ be a valid p-value for the association test at variant $j$, and let $\omega_j$ be its nonnegative weight, with $\sum_{j=1}^{m}\omega_j=1$. The ACAT-V statistic and combined p-value are $$ T_{\mathrm{ACAT-V}}=\sum_{j=1}^{m}\omega_j\tan\!\left[\pi\left(\frac{1}{2}-p_j\right)\right], \qquad p_{\mathrm{ACAT-V}}=\frac{1}{2}-\frac{\arctan(T_{\mathrm{ACAT-V}})}{\pi}. $$ A variant assigned weight zero contributes no evidence and can be omitted.

The input p-values should test the relevant variant-level null hypotheses using a consistent phenotype and covariate specification. They may be two-sided p-values for no association; if directional tests are used, their hypotheses and directions need to be chosen coherently. The combined test concerns the set-level global null that none of the included variants is associated under the specified model. Rejecting that null is evidence against the global null, not proof that every variant is associated.

How the Cauchy Combination Works

If $p_j$ is uniform on $(0,1)$ under its null, then the transformed value $X_j=\tan[\pi(1/2-p_j)]$ has a standard Cauchy distribution. This follows by transforming a uniform variable to an angle uniformly distributed over $(-\pi/2,\pi/2)$, whose tangent has the standard Cauchy distribution. The transformation is decreasing in $p_j$: very small p-values become large positive values, p-values near one-half become values near zero, and p-values above one-half become negative values.

Under independence, a weighted sum of independent standard Cauchy variables is again standard Cauchy when the weights are nonnegative and sum to one. Thus, in that case, the stated upper-tail calculation gives the exact null p-value. Variant tests in genomic data are often dependent because of linkage disequilibrium (LD). With dependence, the standard ACAT-V calculation is generally used as a tail approximation; its accuracy is often strong for small combined p-values, but it is not an exact finite-sample guarantee for every possible dependence structure. Marginal calibration and the dependence pattern both matter.

Justification for the tail behavior: For a small input p-value $p_j$, the transformed value satisfies $$ \tan\!\left[\pi\left(\frac{1}{2}-p_j\right)\right] =\cot(\pi p_j)\approx\frac{1}{\pi p_j}, $$ because $\tan(x)\approx x$ near zero. If one such term dominates the weighted sum, then $T_{\mathrm{ACAT-V}}\approx\omega_j/(\pi p_j)$. For large positive $T_{\mathrm{ACAT-V}}$, the upper-tail p-value is approximately $1/(\pi T_{\mathrm{ACAT-V}})$, so the combined p-value is approximately $p_j/\omega_j$. This approximation explains why an unusually small variant p-value can strongly influence ACAT-V even when many other p-values are unremarkable. It is an explanation of tail behavior, not an exact identity for a general collection of dependent p-values.

Calculating ACAT-V

Worked Example: Combining Four Variant Tests

Suppose a hypothetical set has four variants with p-values $0.25,\ 0.75,\ 0.50,\ 0.01$, and weights $0.10,\ 0.10,\ 0.20,\ 0.60$, respectively. The weights are nonnegative and sum to one. Assume the variant-level p-values have been calibrated for their null hypotheses.

The corresponding Cauchy transforms are $1$, $-1$, $0$, and approximately $31.8205$. Therefore,

T_{\mathrm{ACAT-V}} =(0.10)(1)+(0.10)(-1)+(0.20)(0)+(0.60)(31.8205) =19.0923.

Then

p_{\mathrm{ACAT-V}} =\frac{1}{2}-\frac{\arctan(19.0923)}{\pi} \approx 0.01666.

The set-level result is a combined p-value of about $0.0167$. It is evidence against the global null for this set under the test's calibration assumptions. It is not a 1.67% probability that the null is true. The very small input p-value has substantial influence here because it also has the largest weight.

When Only a Few Variants Carry Evidence

ACAT-V is particularly responsive to sparse evidence: most variants can have p-values near one-half while one or two have very small p-values. The weight assigned to the informative variant matters. With equal weights over many variants, the evidence from one variant is diluted; assigning a larger prespecified weight to a variant class can reduce that dilution, but only if the weighting scheme is scientifically justified and specified appropriately.

Worked Example: One Small P-Value Among Five Variants

Consider a hypothetical five-variant set with one variant p-value of $0.001$, four p-values of $0.50$, and equal weights of $0.20$. The p-values are assumed to be valid under their respective nulls, and the weights sum to one. Since $\tan[\pi(1/2-0.50)]=0$, the four non-extreme variants contribute nothing to the statistic.

For the remaining variant,

\tan[\pi(1/2-0.001)] =\cot(0.001\pi) \approx 318.3099, \qquad T_{\mathrm{ACAT-V}}=(0.20)(318.3099)=63.6620.

The resulting combined p-value is

p_{\mathrm{ACAT-V}} =\frac{1}{2}-\frac{\arctan(63.6620)}{\pi} \approx 0.00500.

The small p-value remains influential, but the combined value is larger than $0.001$, reflecting its weight of $0.20$. This illustrates the approximate $p_j/\omega_j$ behavior for a dominating variant. It also shows why equal weighting over a larger set can dilute a single signal.

Weights can encode prespecified priorities, such as variant annotations or a choice to emphasize a class of variants. The weights must be nonnegative for the standard construction described here and must be normalized to sum to one. A weight is not an estimated effect size, and an increased weight does not establish that a variant is more likely causal. If weights or variant inclusion are selected using the same association results, the nominal combined p-value may fail to account for that selection.

Worked Example: Applying Annotation Weights

Suppose a hypothetical three-variant set uses prespecified annotation weights $0.50,\ 0.30,\ 0.20$, which sum to one. The corresponding calibrated variant p-values are $0.04,\ 0.40,\ 0.50$. Their transformed values are approximately $7.9158,\ 0.3249,\ 0$, respectively. The first transformation is $\cot(0.04\pi)$, and the second is $\tan(0.10\pi)$.

The weighted statistic is

T_{\mathrm{ACAT-V}} =(0.50)(7.9158)+(0.30)(0.3249)+(0.20)(0) =4.0554.

Thus,

p_{\mathrm{ACAT-V}} =\frac{1}{2}-\frac{\arctan(4.0554)}{\pi} \approx 0.07696.

This result is not below a $0.05$ threshold. The weighted p-value reflects the evidence and the declared weights; it is not a weighted average of the three input p-values. The assumptions checked here are that all three p-values target the intended variant-level nulls and that the weights were defined independently of the observed association evidence.

Dependence, Diagnostics, and Failure Modes

The variants need not be independent for ACAT-V to be useful, but LD means the exact independent-Cauchy argument does not apply. A practical analysis should therefore check that each input p-value is calibrated for the sample size, phenotype, covariates, and test used. Sparse binary outcomes can require particular care with calibration; the earlier material on saddlepoint calibration for sparse binary-phenotype tests is relevant when that setting applies. Combining poorly calibrated inputs cannot repair their calibration.

Several checks help identify problems before interpreting a set-level result:

Inspect the input p-values and weights. Confirm that the intended variants are present, weights are finite and nonnegative, and their normalized values sum to one.
Check numerical limits. A computed p-value of exactly zero may be caused by numerical underflow rather than an actual zero probability. Use a method that retains sufficient precision for extreme tails; do not silently replace zero with an arbitrary small number.
Consider LD and repeated information. Highly correlated variants can carry overlapping evidence. ACAT-V's tail approximation is not a license to treat those p-values as independent, and duplicated variants or changed set membership can alter the result and its weighting.
Separate set testing from follow-up. A significant combined p-value does not identify the responsible variant or establish a causal effect. Variant-level follow-up and appropriate replication are separate inferential tasks.
Account for testing many sets. The ACAT-V p-value is for the specified set. If many sets are tested, the broader analysis must address the resulting multiple testing.

How ACAT-V Differs From Other Set Tests

ACAT-V is a p-value-combination procedure, not a burden statistic or a score statistic. Burden tests combine variant data into an aggregate genetic measure and can be sensitive to effects that act in opposing directions. SKAT uses a variance-component score statistic and is designed to accommodate heterogeneous variant effects. ACAT-V instead combines marginal variant-level significance measures through Cauchy transformations; it does not require all variants to have similar effect directions, but it also does not provide a unified set-effect estimate.

ACAT-O, covered in the earlier material on omnibus gene-based testing, combines p-values from multiple set-level tests or choices of analysis. ACAT-V combines p-values across variants within a set. The two methods therefore combine evidence at different levels. Neither a small ACAT-V p-value nor its robust tail behavior eliminates the need for valid inputs, sensible set definition, and appropriate handling of multiple testing.

Common pitfall: “One variant has a tiny p-value, so the ACAT-V p-value must equal that p-value.” This is incorrect. Its contribution is multiplied by its weight, and the other transformed terms can also affect the statistic. When one term dominates, a useful approximation is $p_{\mathrm{ACAT-V}}\approx p_j/\omega_j$, not $p_j$. The combined result must be calculated using the full weighted statistic.

Check Your Understanding

Answer each question, then compare with the solutions below.

Under what condition on the weights is the standard ACAT-V statistic defined here normalized?
Three variants have p-values $0.50,\ 0.50,\ 0.01$ and equal weights. What is the approximate ACAT-V p-value?
Why does a small ACAT-V p-value not show that every variant in the set is associated?
What is the key distinction between ACAT-V and ACAT-O?

Solutions. (1) The weights must be nonnegative and sum to one. (2) The first two transformed values are zero. The third is $\cot(0.01\pi)\approx31.8205$, so $T=(1/3)(31.8205)\approx10.6068$. Hence $p_{\mathrm{ACAT-V}}=1/2-\arctan(10.6068)/\pi\approx0.0299$. (3) The test evaluates evidence against the global null that none of the included variants is associated; it does not establish an alternative for each variant. (4) ACAT-V combines variant-level p-values within a set, whereas ACAT-O combines p-values from multiple set-level analyses or component tests.